Computer observer design for dynamic PET image using time domain filtering

We propose a method to design linear computer observer to evaluate the quality of dynamic PET image. Instead of applying a joint space-time template in the observer design, the proposed suboptimal method separates the temporal and spatial processing. We argue that this framework provides a reasonable way to measure the quality of the dynamic image with moderate complexity. Further more, we study three possible choices for temporal filtering, including temporal matched filter, Patlak analysis and a new method called minimum variance distortionless response (MVDR) filtering. The temporal matched filter sets an upper bound of SNR among all temporal linear filtering methods in terms of SNR, but its output may not be suitable for a physiologically quantitative analysis. Theoretically, the outputs of Patlak and MVDR are inferior to that of the matched filter, but their outputs are physiological related parameters. Monte Carlo simulation and the ROC study are employed to test these three methods. The results show that the MVDR filtering provides the higher area under the curve (AUC) over Patlak method, and its AUC is close to that of temporal matched filter in the high SNR cases, which implies a very good overall observer performance. Therefore, MVDR filtering could be a potential candidate for time domain processing of dynamic image observer in our design.